Cloudpickle and data classes do not play well together

[b78]

When restoring a value of a dataclass with cloudpickle, it breaks:

Example

import dataclasses
from dataclasses import dataclass


@dataclass
class Range:
    start: int
    end: int


v = Range(1, 2)
print("defined")
print(dataclasses.fields(v))

import cloudpickle

print("imported cloudpickle")
print(dataclasses.fields(v))

v_dumped = cloudpickle.dumps(v)

print("dumped with cloudpickle")
print(dataclasses.fields(v))

vv = cloudpickle.loads(v_dumped)
print("loaded with cloudpickle")
print(dataclasses.fields(v))

print("restored value")
print(dataclasses.fields(vv))

Leads to

defined
(Field(name='start',type=<class 'int'>,default=<dataclasses._MISSING_TYPE object at 0x109634ca0>,default_factory=<dataclasses._MISSING_TYPE object at 0x109634ca0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), Field(name='end',type=<class 'int'>,default=<dataclasses._MISSING_TYPE object at 0x109634ca0>,default_factory=<dataclasses._MISSING_TYPE object at 0x109634ca0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
imported cloudpickle
(Field(name='start',type=<class 'int'>,default=<dataclasses._MISSING_TYPE object at 0x109634ca0>,default_factory=<dataclasses._MISSING_TYPE object at 0x109634ca0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), Field(name='end',type=<class 'int'>,default=<dataclasses._MISSING_TYPE object at 0x109634ca0>,default_factory=<dataclasses._MISSING_TYPE object at 0x109634ca0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
dumped with cloudpickle
(Field(name='start',type=<class 'int'>,default=<dataclasses._MISSING_TYPE object at 0x109634ca0>,default_factory=<dataclasses._MISSING_TYPE object at 0x109634ca0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD), Field(name='end',type=<class 'int'>,default=<dataclasses._MISSING_TYPE object at 0x109634ca0>,default_factory=<dataclasses._MISSING_TYPE object at 0x109634ca0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD))
loaded with cloudpickle
()
restored value
()

I.e. as soon as we restore, the type of the data class is destroyed

[b78]

When I move the type definition into it's own file it works.

[drewm1980]

I came here as a user with a different issue, but iirc pickles were never self-describing; they always needed the definitions for non-standard-library types to deserialize correctly. So if you dump the data from a script, the pickle doesn't record a real location from whence the type came, since it was never imported, just defined in the interpreter. Then the deserializer has nothing to import to reconstruct the type.

[rsokl]

Duplicate of #386